On Apr 13, Chris Shoemaker wrote:
=========>
  	Just to get a feel for how difficult/easy classification is I ran
  a commercial decision tree classifier on the connect-4 data.  The results
  are below, but the summary is that it gets a 80.1% accuracy in 9.3 sec
  with a tree of 2595 nodes.
  	I wouldn't set a goal of beating this accuracy.  I would be
  immeasurably impressed/surprised if anyone got this high a classification
  accuracy from their system.  This software is VERY good.  Think of this as
  an upper bound for reasonable results.  So, if you're getting within a few
  percent of this, I'd say you're doing _very_ well.
  	On the other hand, you should view the class distribution as a
  lower bound for reasonable results.  Your system should perform
  somewhat better than a classifer than merely predicts "win" every time
  (65.83%).
  	Good luck,
  	Chris
  
  
  
  
  
  *********
  Evaluation on training data (45038 cases):
  
              Decision Tree
            ----------------
            Size      Errors
  
            2595 5741(12.7%)   <<
  
  
             (a)   (b)   (c)    <-classified as
            ----  ----  ----
           28398   950   272    (a): class win
            1606  9199   284    (b): class loss
            1710   919  1700    (c): class draw
  
  
  Evaluation on test data (22519 cases):
  
              Decision Tree
            ----------------
            Size      Errors
  
            2595 4484(19.9%)   <<
  
  
             (a)   (b)   (c)    <-classified as
            ----  ----  ----
           13632   902   319    (a): class win
            1244  3983   319    (b): class loss
            1057   643   420    (c): class draw
  
  
  Time: 9.3 secs
<=========